phonon-31m English for Compute Flow

phonon-31m English for Compute Flow

Speaks English text through Arm's Compute Flow, from a 31M text-to-speech model packaged for the stock pocket-tts brick. For developers who use Compute Flow and want a small voice.

Made by gradium on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is phonon-31m-en, an English text-to-speech model, packaged for Arm's Compute Flow. It works on the standard tts:pocket-tts brick unchanged, since Phonon shares Pocket TTS's architecture. It is the same model, and only the file format differs.

Who it suits: It suits developers who already use Compute Flow and want a small English voice. Other Phonon runtimes cover the command line, Python, browsers, iOS and Android.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Software | Compute Flow developers | Add a speech step to a pipeline | Speak a short message from a program
Research | Speech researchers | Check a model across several runtimes | Compare the package with the Phonon runtime
Embedded | Teams building small devices | Try a compact voice model | Spoken prompts from a small app

If you do not use Compute Flow, you can skip it. The card names other repositories for llama.cpp and audio.cpp.

Where to use it:

• Arm's Compute Flow, on the stock tts:pocket-tts brick
• A native build made with compute-flow build native
• The tts_cli program that the build produces

Title: Jobs it does
Use case | Who | What they get back
Speak a line of English | Developers | Audio for a prompt you pass in
Offline speech steps | Pipeline builders | A package that runs from a manifest file

How to set it up:

• Compute Flow, with its command line tool
• Python, only if you download the files by hand with the Hugging Face tools
• Disk space for a small model, about 9 MB at 4-bit by Hugging Face's estimate

1. Install the model in Compute Flow once the brick lists it.
2. Build the native program.
3. Run tts_cli with the manifest and your text.

compute-flow model install tts:pocket-tts phonon-31m-en
compute-flow build native
./build_native/tts_cli -m @manifest/pocket-tts/phonon-31m-en/phonon_manifest.json \
  -p "Hello from Phonon." -t 1

hf download gradium/phonon-31m-en-compute-flow --local-dir phonon-31m-en
(cd phonon-31m-en && shasum -a 256 -c SHA256SUMS)
./build_native/tts_cli -m phonon-31m-en/phonon_manifest.json -p "Hello from Phonon." -t 1

How to use it: Run tts_cli with the manifest path and the text to speak. If the brick does not list the model yet, download this repository, check the checksums, and pass the manifest's path. Spell out numbers, symbols and abbreviations in words first.

Title: Commands in the quickstart
Tool | What it does
compute-flow model install | Adds the model to the pocket-tts brick
compute-flow build native | Builds the native program
tts_cli | Speaks the text given with the p option
shasum -a 256 -c SHA256SUMS | Checks the downloaded files against their checksums

Example: You run tts_cli with the manifest and the text "Hello from Phonon." The program speaks that sentence in the package's default voice. The card does not say which output file or device you get.

For newcomers and experts: If you are new, text-to-speech turns written words into a spoken voice. The first small thing to try is the card's quickstart with its one sample sentence.

If you know the field, the package splits the model into ONNX parts, a Q8_0 flow-matching transformer for llama.cpp and an f32 LiteRT Mimi decoder. The card lists four differences from the Phonon runtime, such as an exact versus approximate GELU, and says the audio is not sample-identical. The trade-off is a later first audio, since the voice prompt is fed in again on every request.

Install the model into the brick
Build the native program
Run tts_cli with a manifest and text
Speech is generated frame by frame

Good to know:

• Voice: the package holds the default voice only.
• Length: under the manifest, each request produces at most 500 frames, about 40 seconds of speech.
• Sampling: temperature is 0.3 as in the Phonon runtime, and the model's own end-of-speech threshold is +4.0.
• Differences: the card lists four, including a small shift in the generated latents.
• License: apache-2.0.
• Cost to run: no API key, account or paid service is named.
• Data it reaches: you download files from Hugging Face, and the card names no other access.
• Not for: changing the speaker count or duration settings, which are fixed at their defaults.

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: litert
• Parameters: 18.9M
• Weights: about 9 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 56
• Likes: 3
• License: apache-2.0
• Languages: en
• Based on: gradium/phonon-31m-en
• Tags: litert, tflite, onnx, gguf, text-to-speech, phonon, compute-flow, sentencepiece

Links

• Model: https://huggingface.co/gradium/phonon-31m-en-compute-flow
• Files: https://huggingface.co/gradium/phonon-31m-en-compute-flow/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download gradium/phonon-31m-en-compute-flow

Then follow the usage in its model card. Nothing in this kit runs by itself.